2026.11.05 · 4 min · ai · css · tooling · news
Lint your UI for AI slop.
npx impeccable detect is a linter for design: 62 deterministic rules, no LLM, no API key, each finding tied to a rule and a line. What it found on a card I made sloppy on purpose, what a source scan can't see, and the new --github flag for pull request annotations. As of 2026.11.
AI-built pages share defaults: gradient text, a thick coloured stripe down one side of a card, Inter. Once you've seen them you can't unsee them, but if you're new to design you may not see them at all. A linter can. npx impeccable detect src/ scans your UI code and names each pattern with its rule and line, before you ship.
Watch the 60-second version on Instagram.
What it is
Impeccable's detector is "62 deterministic rules... No LLM, no API key", in its README's words. It's open source (Apache-2.0, by Paul Bakaus) and needs Node 18.19 or later.
npx impeccable detect src/ # scan a folder
npx impeccable detect http://localhost:3000 # scan the running dev serverThe full usage is impeccable detect [options] [file-or-dir-or-url...]. It reads JSX, TSX, Vue, Svelte, Astro, stylesheets and HTML (with its linked local CSS). URLs open in a browser. It's for the web only, not native apps.
What it found on a sloppy card
I wrote a throwaway Hero.tsx and hero.css, deliberately sloppy, and ran impeccable@4.8.0 on them. This is a test sample, not a real project.
The figure is a drawing of the card, so the gradient is hatching and there's no purple, but the four findings and their lines are the real output. Here it is, with the one-line explanation the tool prints under each finding trimmed:
2 files scanned with regex matching only (.css, .tsx).
Rules that need a parsed page (tiny-text, low-contrast and other element rules) did not run on them; findings are an undercount, not a clean bill of health.
For the full check, scan the running dev server (npx impeccable detect http://localhost:PORT) or a static site's built .html.
src/Hero.tsx
line 7: [side-tab] border-l-4
line 6: [gradient-text] bg-clip-text + bg-gradient
line 6: [ai-color-palette] from-purple-500 gradient
src/hero.css (imported by Hero.tsx)
line 2: [overused-font] font-family: Inter
4 anti-patterns found.Read the first three lines. On source files, the rules that need a parsed page, such as tiny-text and low-contrast, don't run, and the tool says so itself. For those, point it at the dev server URL or the built HTML. Text findings go to stderr; --json goes to stdout.
The plain-CSS version from the docs is the classic gradient headline, reported as gradient-text on line 3:
h1 {
background: linear-gradient(90deg, #ff3399, #8833ff);
background-clip: text;
-webkit-background-clip: text;
color: transparent;
}Exit codes, for CI
| Exit code | Means |
|---|---|
| 0 | No primary findings |
| 2 | Primary findings (my sample exited 2) |
| 1 | A target couldn't be scanned |
Advisories never fail a scan.
New in v4.8.0: findings on your pull request
detect itself isn't new; it dates from v2.0 in April 2026. What v4.8.0 (2026.10.26) added is --github, which prints one ::warning line per finding, so each one shows up as an annotation on its line in the pull request. The same release added --list-rules, a digest-checked engine binary, and downgraded crisp 4px+ shadows to an advisory.
The terminal lines are real output on my sample (wrapped, and trimmed after the third line). The pull request is drawn, not a screenshot.
What people get wrong
- "It detects AI-written code." It doesn't check authorship. It flags patterns, and some of its rules are plain quality checks, such as skipped heading levels. In the project's own words, "a finding is a reason to look closer".
- "A clean scan means good design." A static scan doesn't run your JS or see the painted layout, and on source files the element rules don't run at all. One developer also reports five false warnings from links inside a closed
<details>. - "Every finding has to be fixed." Some are deliberate brand choices. Waive them with a reason:
npx impeccable ignores add-value overused-font Inter --reason "Brand font"That last one applies to this site. Its headings are set in Space Grotesk, which is on the current overused-font list next to Inter, Roboto, Fraunces, Geist and Plus Jakarta Sans. The slop catalog also lists decorative grid backgrounds, and I use one behind page headers. I haven't run the detector on this site, so I'm not claiming it's clean.
On a client project I'd run it against the dev server URL in CI with --github, and waive the brand font with an ignore and a reason rather than change fonts to please a linter.
The short version
npx impeccable detect src/is a linter for design: 62 deterministic rules, no LLM, no API key.- Each finding names a rule and usually a line:
side-tab,gradient-text,ai-color-palette,overused-fonton my sample. - A source scan undercounts; scan the dev server URL or built HTML for element rules like
low-contrast. - A finding is a reason to look closer, not proof AI wrote the code. Waive deliberate choices with
ignores add-value. - Exit code 2 means findings; v4.8.0's
--githubturns them into pull request annotations.
Sources: Impeccable changelog · CLI v4.8.0 release · Detector docs · README · Slop catalog · cli.rs · github_annotations.rs · Config docs · npm metadata · Mager, Diving into Impeccable